Post-processing Technology Of Surveillance Image/Video

Outdoor video systems are applied widely and variously in currentsociety. However, as surveillance videos are taken all day long, there arehuge amount of data, which are the biggest challenge we face.On one hand, dynamic weather conditions such as rain and snow canresult in inaccuracy of motion estimation, thus enlarging motioncompensation residual. However, dynamic rain and snow are not usefulinformation for surveillance videos. If we can remove rain and snow whentaking videos, we could reduce compression disturbing, as a result,improving data storage efficiency.On the other hand, in order to reduce data, image and video are usuallycompressed by different standards. In these standards, image blocks are firsttransformed into a different domain and then quantized subjected to someparameters. Quantization is a lossy process, resulting in details lost. Thisinformation loss is especially obvious when images are heavily compressed.Image degradation appears as blocking artifacts around block boundary orringing artifacts near strong edges. As a result, image deblocking algorithmcan improve storage and searching efficiency and improve subjectivehuman vision feelings.Considering the wide application of surveillance system, videorain-removal and image deblocking algorithm can bring in significance forcurrent society.In order to reduce the disturbing from rainy weathers, we applyanisotropic diffusion tensor as the math foundation. In recent years, theidea of anisotropic nonlinear diffusion(AND) has gained a wide range ofapplications in image noise reducing and image segmentation. Thefundamental idea of this algorithm is to solve the partial difference equation(PDE) with the initial solution based on the original image. It can effectivelysmooth the image while preserving the edges, hence remove the noise. Dueto difficulties introduced by dynamic characteristics of the rain, we chooserain removal as a special case of removing speckle noise. The results show that this filter is very effective in removing rain streaks of different width inimages. Compared to traditional noise-filtering methods, AND does not usean algorithm to detect raindrop trajectory, and instead remove the rainstreaks with a self-adapting filter, therefore achieving good robustness andcomputing speed.In order to reduce quantization parameter, properties of thequantization noise are first reviewed in this paper. Then, nonlocal meansfilter is selected to remove noise, which produces satisfied dequantizationresults with less resource requirements. This filter is adjusted adaptivelywith image content and noise distribution which are described by twocharacteristics. Linear regression is used to determine the relationship of theoptimal filter parameter and the two characteristics. Experimental resultsindicate that the proposed algorithm gets better denoising performance thanthe peer ones on the aspects of image quality, resource requirement andexecution speed.